search

martinkersner / train-DeepLab

Train DeepLab for Semantic Image Segmentation
MIT License
172 stars 76 forks source link

Wrong prediction results? #7

Closed koizora223 closed 8 years ago

koizora223 commented 8 years ago

Hi @martinkersner I run deeplab.py using the downloaded official caffemodel(train2_iter_8000.caffemodel). But I get wrong results.The object seems to locate in the wrong position. I have checked the output of each layer and found the output of fully connected layer fc6 incorrect. There are my results.Please give me some advices.Thanks!! c-label

a-label

Chen

martinkersner commented 8 years ago

Hi @koizora223,

What do you mean by "found [that] the output of fully connected layer fc6 [is] incorrect"? How did you find it out?

Cheers, Martin

koizora223 commented 8 years ago

Hi @martinkersner I have showed the output of 'pool5a' and ' fc6' layer using the code below in segmenter.py. feat = self.blobs['pool5a'].data[0, :21] vis_square(feat) feat = self.blobs['fc6'].data[0, :21] vis_square(feat) And I got the following results:The coarse plane output from pooling layer(row2,col1) is at the right position,but wrong after fc6 layer(row1,col 1). The final prediction result is at the wrong position as the second image. What cause this? figure_0 figure_1

Thanks, Chen

martinkersner commented 8 years ago

@koizora223 I am sorry, but I don't know.

koizora223 commented 8 years ago

Hi @martinkersner It is resovled by just compiling without cuDNN enabled.Thanks all the same! And I have another question .Do you use cuDNN for training?

Cheers!

martinkersner commented 8 years ago

@koizora223 Great! I haven't used cuDNN for either training or testing.